Deep Learning and Feature Engineering for Solar Flare Prediction
Abstract
Space weather events can inflict chaos on our modern, technologically-dependent world, such as degradation, and sometimes even disruption, for our telecommunications, power grids, navigation systems, and satellite operations. Therefore, improving the accuracy in predicting such events is of the upmost importance. With this in mind, we built a deep learning system capable of predicting M and X class solar flares in a 24 hour time window, and we achieved a True Skill Score of 0.9. The inputs to our deep learning algorithm are magnetograms from the Solar Dynamics Observatory Helioseismic and Magnetic Imager (SDO/HMI), and the output is a binary classification indicating whether or not a solar flare will occur in the time window. The highest True Skill Scores were attained with the convolutional neural networks AlexNet and VGGNet. We are currently working to incorporate SDO Extreme ultraviolet Variability (EVE) time series data into our existing deep learning framework. As we add more data, we anticipate a richer model with increased predictive power. Using examples and results from our convolutional neural network and multilayer perceptron models, we’ll present how we approached challenges and sought to fix them and what methods we propose as we further develop our models. We’ll also present the importance of feature engineering as it relates to solar flare prediction, and the obstacles and rewards of creating our own in-house features, including polarity inversion lines.
- Publication:
-
Solar Heliospheric and INterplanetary Environment (SHINE 2019)
- Pub Date:
- May 2019
- Bibcode:
- 2019shin.confE.156C